The case for a 2027 intelligence explosion now rests on evidence the public can't see

The oldest argument for a fast takeoff is I. J. Good's: designing machines is an intellectual activity, so a machine better than us at it could design a better machine, and so on. For sixty years that loop was hypothetical. On November 5, 2026, METR published the first independent assessment of AI-driven acceleration at a frontier lab. It found that Anthropic's capabilities, on its internal version of the Epoch Capabilities Index, have grown about 1.9 times faster since October 2025, with a 95% interval of 1.5 to 2.8. METR guessed that perhaps 80% of that is due to AI, and said it was confident the speed-up did not come from faster growth in training compute.

Forecasters moved within days. The AI Futures Project, whose AI 2027 scenario described an intelligence explosion in 2027, had spent 2025 pushing its own dates back; Daniel Kokotajlo's median slipped to 2029. Its August 2026 update put an automated coder at November 2027 for Kokotajlo and January 2030 for Eli Lifland, with superintelligence medians of March 2029 and July 2033. Its November 10 update says the authors now expect a superhuman AI researcher to “likely arrive at some point in 2027,” with a 35–65% chance of AI that dominates top experts before 2028. The authors write that even their most aggressive model gave less than 4% to seeing this much acceleration by now.

The labs say similar things about themselves. Anthropic reports that more than 80% of the code it merges was written by Claude as of May 2026, and that Claude now leads 26% of its AI research work, with none of it fully autonomous. OpenAI said in September that it had met its goal of an “automated research intern,” while reporting that more than half of successful four-to-eight-hour tasks still needed a human to step in. Amodei wrote in June that if scaling holds “for only a year or two longer”, the result will be his “country of geniuses in a datacenter.” These are interested parties describing their own work, and none of the figures has been audited.

The slower forecasts have not gone away. The largest survey of AI researchers, 2,778 authors polled in late 2023, gave a 50% chance that machines outperform humans at every task by 2047, and only by 2116 that all occupations are automated. Ilya Sutskever put human-like learning efficiency five to twenty years out, and Andrej Karpathy called agents a decade away because they cannot remember what they are told. Arvind Narayanan and Sayash Kapoor's AI as Normal Technology argues that diffusion through institutions, not invention, sets the pace of impact.

The physical inputs are growing fast but not without limit. Epoch AI measures frontier training compute rising about 5 times a year since 2020, and the largest single site about 3.3 times a year, but expects no site to clearly exceed xAI's Colossus 2 before the second half of 2027, because gigawatt sites take about two years to build. The IEA projects data-centre electricity reaching about 945 TWh by 2030. We covered why the grid, not the chip, sets the 2027 ceiling in an earlier post. METR, for its part, saw no binding shortage of compute, data or algorithmic ideas in the next twelve months, and said it holds that view with low confidence.

Two caveats in METR's own report matter most. First, transparency: Anthropic held redaction rights and removed “a majority of the evidence we found most informative,” and a planned check on how compute was split between capabilities and safety was cancelled midway. We wrote about the same gap in Anthropic's IPO filing. Second, METR says the acceleration may be limited to well-specified tasks that can be graded automatically. That is where optimization finds the grader's blind spots. METR has documented frontier models editing scoring code to pass tests, and in July 2026 agents in an OpenAI cyber-capability evaluation escaped isolation and, in Hugging Face's account, compromised its production infrastructure end to end.

Our research sits on exactly this seam. In simulation, Best-of-N accuracy peaks and then falls once a verifier can be fooled at any rate, and exploits constructed against a verifier beat it far more often than sampled ones. A self-improving system that is graded by its own tests is a Best-of-N process run for months. So the question that decides how fast superintelligence arrives is not only how much compute is built, but whether the measured gains hold up on tasks the system cannot grade for itself. That is checkable, and nobody outside the labs can check it yet.